Applications of signal processors SIGNAL GENERATION on digital signal processors Author: Grzegorz Szwoch Gdańsk University of Technology, Department of Multimedia Systems Introduction ▪ Usually, signal processors operate on signals that are fed to its inputs. ▪ We can also use DSPs to generate signals. ▪ In this lecture, we will talk about: • generating digital harmonic signals, • generating a sine wave, • generating pseudo-random signals (noise), • generating signals by reading samples from memory, • interpolation of samples stored in memory. Sawtooth wave ▪ Harmonic signals: their spectrum consists of partials at harmonic frequencies – multiples of the fundamental frequency. ▪ Example: sawtooth wave. Time and spectral plot: Sawtooth wave ▪ Amplitude changes linearly. ▪ To generate the wave, we use an accumulator – we sum up the consecutive amplitude steps. ▪ Initialization: int amplitude = 0; const int step = ???; ▪ For each sample, output y: y = amplitude; amplitude = amplituda + step; ▪ What is the value of step? Calculating the amplitude step ▪ Let’s assume frequency 1 Hz (period 1 s), fs = 48 kHz. ▪ We need 48000 samples to change amplitude from -32768 to 32768. ▪ Amplitude change per one sample is: 232768 d = = 1.365333... 48000 ▪ And if we need 100 Hz (sample 0.01 s)? 232768 d = = 136.5333... 48000 /100 Calculating the amplitude step ▪ For any frequency f, amplitude step as a Q15 number is: d = round( f * 1.36533) ▪ For example: f = 440 Hz → d = 601 ▪ If we need to compute this step in code: 65536 2 22368 d = f = f = ( f *22368 ) 14 48000 32768 ▪ Remember that it’s not possible to write any frequency value in a fixed-point notation. Overflow in sawtooth generation ▪ Important: range overflow occurs when amplitude steps are accumulated, for example: 32750 + 25 = „32775” = –32761 ▪ Amplitude “wraps around” - this is exactly what we need! ▪ It is one of rare cases in which range overflow is actually useful. Square / pulse wave ▪ Another harmonic signal: square or pulse wave. ▪ Signal amplitude changes between -A and +A. ▪ Pulse width: a ratio of duration of the positive part to the wave period (0 to 1). ▪ Time and spectral plots for pulse width = 0.5: Square / pulse wave ▪ Square wave may be calculated from the sawtooth wave: if (amplitude < threshold) y = 32767; // or another amplitude else y = -32768; ▪ Value of threshold depends on the pulse width: threshold = 2 * pulse_width – 1 ▪ A regular square wave (50/50): threshold = 0. Aliasing problem ▪ Analog harmonic waves (such as square or sawtooth) have infinite spectrum. ▪ If we try to generate these waves digitally “from definition”, usually aliasing will occur. ▪ The problem increases for higher wave frequencies. ▪ The resulting signal is inharmonic. Aliasing problem There ware various method of generating alias-free waves. ▪ Generating waves with higher sampling frequency (oversampling), then decimation. ▪ Using Fourier series. For example, sawtooth: N A A k sin(2knf / fs) x(n) = − (−1) 2 k =1 k • for kf below Nyquist frequency, • the signal is distorted in time domain (lack of high frequency components). Sine generation ▪ Phase of a sine wave looks like a sawtooth wave. ▪ We know how to generate a sawtooth wave. Now, we have to convert phase (angle) into the amplitude. ▪ We can approximate the sine with Taylor series: x3 x5 x7 x9 sin(x) x − + − + −... 3! 5! 7! 9! ▪ Function sine from DSPLIB for C55x uses this method. Sine wave generation with DSPLIB ▪ x – pointer to the buffer with phase (angle) values. We generate a sawtooth wave of a desired frequency and write its values to the buffer. ▪ r – pointer to the buffer in which values of a sine wave will be stored. ▪ nx – number of samples to generate (buffer length). Note that sine function does not generate a sine wave signal by itself, it only computes sine values from given angles. Sine wave from IIR generator ▪ Alternative method of sine wave generation: we use a marginally-stable second-order IIR system. ▪ We use an impulse to start the generator: y(0) = –sin(2f/fs), y(1) = 0 ▪ The system goes into oscillations – generates sine wave values. ▪ On a fixed-point DSP, implementation is problematic (insufficient numerical precision). 2 f a = 2cos y(n) = a y(n −1) − y(n − 2) fs White noise generation ▪ White noise – a random signal with flat spectrum. ▪ To generate a digital noise, we use (pseudo)random number generators – RNG. ▪ Noise samples are computed by the algorithm. ▪ Example of a simple noise generation algorithm: LCG – linear congruent generator: y(n) = a y(n −1)+ b mod M mod – modulo, remainder of integer division by M ▪ For professional applications, such as cryptography, more accurate algorithms are needed (e.g., Mersenne Twister). White noise generation ▪ Initial value y(0) is called a seed. Given the same seed, the algorithm will always generate the same sequence of pseudo-random numbers. ▪ In practice, we set the seed to a constantly changing value, e.g., a counter of processor cycles. ▪ Example: a = 2045, b= 0, M = 220, y(0) = 12345. Time and spectrum plots: White noise generation in DSPLIB ▪ Initialization – just once, when the program starts: rand16init(); ▪ Writing nr samples into buffer r: LCG: a = 31821, b = 13849, M = 65536 rand16(bufor, 2048); Signal from a wavetable ▪ Any signal can be generated by reading its samples stored in memory, in a wavetable (a buffer of samples). ▪ For example, we can store 480 samples of one period of a sine wave. ▪ If we read them with speed of 48 kHz, we get a sine wave of frequency 100 Hz. ▪ If we read every second sample, we have 240 samples per period, therefore f = 48000 / 240 = 200 Hz ▪ If we loop the samples we read, we get a continuous wave. ▪ The problem: how can we generate any frequency? Signal from a wavetable ▪ A general case: generating wave of any frequency. ▪ Step to move the read index in the wavetable: s = f N / fs (N – number of samples in wavetable). ▪ For f = 456 Hz, N = 4800: s = 45.6 ▪ Usually, step s is not an integer. ▪ So, we need to “read between samples”. ▪ Interpolation of samples: estimating values between samples stored in the memory. Linear interpolation ▪ The simplest interpolation is linear. We connect the known samples with a straight line, and we look for a value at a given position on the line. ▪ Let index = 45.6. We interpolate between samples x[45] and x[46] – the previous and the next one. A2 − A1 A − A1 = n2 – n1 = 1 A2 n2 − n1 n − n1 A A = A1 + (A2 − A1 )(n − n1 ) A1 n1 n n2 x45.6 = x45+ (x46− x45)0.6 Reading samples with interpolation // Example: index = 45.6 int index_c = 45; // integer part (45), int int index_u = 19661; // fractional part (0.6), Q15 // Read samples from table int a1 = buffer[index_c]; // the previous sample int a2 = buffer[index_c+1]; // the next sample // linear interpolation long a = index_u * (long)(a2 – a1) // (n-n1)*(a2-a1) a = (_sround(a<<1) >> 16) + a1 // + a1 // value of the interpolated sample y = (int)a; Other interpolation methods ▪ Linear interpolation is simple, but not accurate. ▪ More accurate methods: • polynomial interpolation – of degree 2 (square), 3 (cubic) and higher degrees. • interpolation with sin(x)/x functions (sinc interpolation): Signal from a wavetable ▪ The more samples in the buffer, the better (lower interpolation errors). ▪ We can store any signal in the wavetable. ▪ This methods works fine if we read and loop a wave period. ▪ If we do not loop, frequency change results in changing the duration – the signal becomes shorter or longer. ▪ Interpolation of signals with complex spectrum, such as square wave, may result in aliasing. Usually, we need to store a number of wave versions of different frequencies in the wavetable. Signal from a wavetable Reading a sound signal from table with different steps: ▪ step = 1: • samples read with the original sampling frequency, • sound pitch – the same as the original sampled sound; ▪ step < 1: • we read samples more slowly – sound becomes longer • the sound pitch is lower than the original; ▪ step > 1: • we read samples quicker – sound becomes shorter • the sound pitch is higher than the original. Sampler ▪ A practical example: a sampler – digital musical instrument that plays back sound samples stored in memory. Only fragments of samples are looped, or they are not looped. ▪ A DSP in the sampler does transposition - changes the pitch of generated sounds by altering the step of memory read index and by interpolation. ▪ Temporal distortions occur – the sound becomes longer or shorter. Therefore, we need to use a set of samples with different pitch (multisampling)..